Model generation apparatus, estimation apparatus, model generation method, and computer-readable storage medium storing a model generation program including training a model by converting training data to deliberately reduce estimation performance for improved defect detection and feature identification

Abstract

A model generation apparatus according to one or more embodiments executes operations, with respect to each of learning data sets. The operations includes training a second estimator so that an estimation result obtained from a second estimator conforms to second correct answer data; training a coder so that an estimation result obtained from the second estimator does not conform to the second correct answer data; and training the coder and the first estimator so that an estimation result obtained from a first estimator conforms to first correct answer data. The model generation apparatus executes operation of the training the second estimator and the training the coder alternately and repeatedly.

The invention claimed is: 
     
       1. A model generation apparatus comprising a processor coupled to a memory storing a program and configured with the program to perform operations comprising:
 acquiring a plurality of learning data sets each constituted by a combination of training data, first correct answer data that indicates a first feature comprised in the training data, and second correct answer data that indicates a second feature comprised in the training data and is different from the first feature, the training data comprises image data; and 
 executing machine learning of a learning model that comprises a coder, a first estimator, and a second estimator, wherein the coder is configured to convert received input data into a feature amount, the first estimator is configured to accept input of an output value of the coder, and estimate a first feature comprised in the input data from the feature amount, and the second estimator is configured to accept input of the output value of the coder, and estimate a second feature comprised in the input data from the feature amount, wherein 
 the processor is configured with the program to perform operations such that executing the machine learning of a learning model comprises:
 training the second estimator so that, with respect to each of the learning data sets, an estimation result obtained from the second estimator by giving the training data to the coder conforms to the second correct answer data; 
 training the coder so that, with respect to each of the learning data sets, an estimation result obtained from the second estimator by giving the training data to the coder does not conform to the second correct answer data by converting the training data into the feature amounts such that estimation performance of the second estimator is deliberately reduced; 
 training the coder and the first estimator so that, with respect to each of the learning data sets, an estimation result obtained from the first estimator by giving the training data to the coder conforms to the first correct answer data, and 
 executing adversarial learning in the coder and the second estimator by executing training in the second estimator, and executing training in the coder that has been trained such that the estimation result obtained from the second estimator does not conform to the second correct answer data and estimation performance of the second estimator is deliberately reduced, alternately and repeatedly, and 
 
 training the coder so that, with respect to each of the learning data sets, the estimation result obtained from the second estimator by giving the training data to the coder does not conform to the second correct answer data such that estimation performance of the second estimator is deliberately reduced comprises:
 training the coder using dummy data that corresponds to the second correct answer data; or 
 training the coder using reverse gradients. 
 
 
     
     
       2. The model generation apparatus according to  claim 1 ,
 wherein the processor is configured with the program to perform operations such that training the coder and the first estimator is repeatedly executed together with training the second estimator and training the coder. 
 
     
     
       3. The model generation apparatus according to  claim 1 ,
 wherein the processor is configured with the program to perform operations such that training the coder comprises
 acquiring, with respect to each of the learning data sets, the dummy data that corresponds to the second correct answer data and is constituted by values different from values of the corresponding second correct answer data, and 
 training the coder so that the estimation result does not conform to the second correct answer data by training the coder so that the estimation result obtained from the second estimator by giving the training data to the coder conforms to the dummy data. 
 
 
     
     
       4. The model generation apparatus according to  claim 3 ,
 wherein the dummy data comprises the second correct answer data of a learning data set that is different from the corresponding learning data set. 
 
     
     
       5. The model generation apparatus according to  claim 1 ,
 wherein the first feature relates to a first component to be subjected to a predetermined estimation, and 
 the second feature relates to a second component that is different from the first component, and affects the predetermined estimation regarding the first component. 
 
     
     
       6. The model generation apparatus according to  claim 1 ,
 wherein the training data comprises image data comprising a foreground and a background, 
 the first feature relates to the foreground, and 
 the second feature relates to the background. 
 
     
     
       7. The model generation apparatus according to  claim 1 ,
 wherein the training data comprises image data in which an object appears, 
 the first feature comprises an attribute of the object, and 
 the second feature is another feature different from the attribute of the object. 
 
     
     
       8. The model generation apparatus according to  claim 7 ,
 wherein the object comprises a product, and 
 the attribute of the object relates to a defect of the product. 
 
     
     
       9. An estimation apparatus comprising a processor coupled to a memory storing a program and configured with the program to perform operations comprising:
 acquiring object data; 
 estimating a first feature comprised in the acquired object data, using the trained coder and the trained first estimator that are generated by the model generation apparatus according to  claim 1 ; and 
 outputting information relating to a result of the estimation of the first feature. 
 
     
     
       10. A model generation method in which a computer performs operations comprising:
 acquiring a plurality of learning data sets each constituted by a combination of training data, first correct answer data that indicates a first feature comprised in the training data, and second correct answer data that indicates a second feature comprised in the training data and is different from the first feature; and 
 executing machine learning of a learning model that comprises a coder, a first estimator, and a second estimator, wherein the coder is configured to convert received input data into a feature amount, the first estimator is configured to accept input of an output value of the coder, and estimate a first feature comprised in the input data from the feature amount, and the second estimator is configured to accept input of the output value of the coder, and estimate a second feature comprised in the input data from the feature amount, wherein 
 executing the machine learning comprises:
 training the second estimator so that, with respect to each of the learning data sets, an estimation result obtained from the second estimator by giving the training data to the coder conforms to the second correct answer data; 
 training the coder so that, with respect to each of the learning data sets, an estimation result obtained from the second estimator by giving the training data to the coder does not conform to the second correct answer data by converting the training data into the feature amounts such that estimation performance of the second estimator is deliberately reduced; 
 training the coder and the first estimator so that, with respect to each of the learning data sets, an estimation result obtained from the first estimator by giving the training data to the coder conforms to the first correct answer data, and 
 executing adversarial learning in the coder and the second estimator by executing training in the second estimator, and executing training in the coder that has been trained such that the estimation result obtained from the second estimator does not conform to the second correct answer data and estimation performance of the second estimator is deliberately reduced, alternately and repeatedly, and 
 
 training the coder so that, with respect to each of the learning data sets, the estimation result obtained from the second estimator by giving the training data to the coder does not conform to the second correct answer data such that estimation performance of the second estimator is deliberately reduced comprises:
 training the coder using dummy data that corresponds to the second correct answer data; or 
 training the coder using reverse gradients. 
 
 
     
     
       11. A non-transitory computer-readable storage medium storing a model generation program, which when read and executed, for causing a computer to perform operations comprising:
 acquiring a plurality of learning data sets each constituted by a combination of training data, first correct answer data that indicates a first feature comprised in the training data, and second correct answer data that indicates a second feature comprised in the training data and is different from the first feature; and 
 executing machine learning of a learning model that comprises a coder, a first estimator, and a second estimator, wherein the coder is configured to convert received input data into a feature amount, the first estimator is configured to accept input of an output value of the coder, and estimate a first feature comprised in the input data from the feature amount, and the second estimator is configured to accept input of the output value of the coder, and estimate a second feature comprised in the input data from the feature amount, wherein 
 executing the machine learning comprises:
 training the second estimator so that, with respect to each of the learning data sets, an estimation result obtained from the second estimator by giving the training data to the coder conforms to the second correct answer data; 
 training the coder so that, with respect to each of the learning data sets, an estimation result obtained from the second estimator by giving the training data to the coder does not conform to the second correct answer data by converting the training data into the feature amounts such that estimation performance of the second estimator is deliberately reduced; 
 training the coder and the first estimator so that, with respect to each of the learning data sets, an estimation result obtained from the first estimator by giving the training data to the coder conforms to the first correct answer data, and 
 executing adversarial learning in the coder and the second estimator by executing training in the second estimator, and executing training in the coder that has been trained such that the estimation result obtained from the second estimator does not conform to the second correct answer data and estimation performance of the second estimator is deliberately reduced, alternately and repeatedly, and 
 
 training the coder so that, with respect to each of the learning data sets, the estimation result obtained from the second estimator by giving the training data to the coder does not conform to the second correct answer data such that estimation performance of the second estimator is deliberately reduced comprises:
 training the coder using dummy data that corresponds to the second correct answer data; or 
 training the coder using reverse gradients. 
 
 
     
     
       12. The model generation apparatus according to  claim 2 ,
 wherein the processor is configured with the program to perform operations such that training the coder comprises
 acquiring, with respect to each of the learning data sets, the dummy data that corresponds to the second correct answer data and is constituted by values different from values of the corresponding second correct answer data, and 
 training the coder so that the estimation result does not conform to the second correct answer data by training the coder so that the estimation result obtained from the second estimator by giving the training data to the coder conforms to the dummy data. 
 
 
     
     
       13. The model generation apparatus according to  claim 2 ,
 wherein the first feature relates to a first component to be subjected to a predetermined estimation, and 
 the second feature relates to a second component that is different from the first component, and affects the predetermined estimation regarding the first component. 
 
     
     
       14. The model generation apparatus according to  claim 3 ,
 wherein the first feature relates to a first component to be subjected to a predetermined estimation, and 
 the second feature relates to a second component that is different from the first component, and affects the predetermined estimation regarding the first component. 
 
     
     
       15. The model generation apparatus according to  claim 4 ,
 wherein the first feature relates to a first component to be subjected to a predetermined estimation, and 
 the second feature relates to a second component that is different from the first component, and affects the predetermined estimation regarding the first component. 
 
     
     
       16. The model generation apparatus according to  claim 2 ,
 wherein the training data comprises image data comprising a foreground and a background, 
 the first feature relates to the foreground, and 
 the second feature relates to the background. 
 
     
     
       17. The model generation apparatus according to  claim 3 ,
 wherein the training data comprises image data comprising a foreground and a background, 
 the first feature relates to the foreground, and 
 the second feature relates to the background. 
 
     
     
       18. The model generation apparatus according to  claim 4 ,
 wherein the training data comprises image data comprising a foreground and a background, 
 the first feature relates to the foreground, and 
 the second feature relates to the background. 
 
     
     
       19. The model generation apparatus according to  claim 2 ,
 wherein the training data comprises image data in which an object appears, 
 the first feature comprises an attribute of the object, and 
 the second feature is another feature different from the attribute of the object. 
 
     
     
       20. The model generation apparatus according to  claim 3 ,
 wherein the training data comprises image data in which an object appears, 
 the first feature comprises an attribute of the object, and 
 the second feature is another feature different from the attribute of the object.